What an OpenAI Select Partner can do for you
We are an OpenAI Select Partner. OpenAI looked at what we build and how we run it, and brought us into the partner network. Here is what that changes about the work we can do on your project.
Where the model runs is a decision, not a default
Being an OpenAI partner does not mean every job ends with a call to the OpenAI API. It means knowing the platform well enough to say when it is the right answer and when it is not.
The API is the fastest route and usually the cheapest way to get something live. Where a client is already on AWS, or needs the model running inside their own account under their own controls, we use Bedrock. It carries OpenAI's open-weight models alongside Anthropic, Meta and Mistral behind one interface, so changing model later does not mean rewriting the application.
We ship on AWS. That is the infrastructure we know properly, and it is where we would put anything that has to sit close to your own data.
Most clients are fine on the API. The ones who are not are usually in a regulated setting, holding somebody else's data under contract, or doing enough volume that the economics change. Those are different builds, and that conversation is better had before the architecture is set than after.
What we build
Agents inside the systems you already run. Reading an enquiry and drafting the quote against your live rate card. Chasing overdue invoices on your terms. Triaging an inbox. Solving a schedule around skills, travel and promises already made. A person approves anything that leaves the business, and every action is logged.
MCP connections. The model never gets your database. It gets a defined set of tools, each with its own permissions, each logged when used. Look up a customer, check stock, draft a document. It cannot do the thing nobody authorised.
Documents and photographs. Delivery notes, invoices, certificates, handwritten site paperwork. Pulling the fields out and putting them where they belong instead of someone retyping them. Vision models on site photos for condition, damage and compliance checks.
Voice. Taking a booking, answering the common questions, capturing a job out of hours, and handing the awkward ones to a person with the context already gathered.
Retrieval over your own material. Specs, drawings, contracts, past jobs. Answers that cite the document they came from, so somebody can check.
Fine tuning, when it earns it. Usually it does not. Most problems people bring us as fine tuning turn out to be retrieval or a better prompt, and we will say so.
Proving it works before it touches a customer
This is the part that separates a demo from a system.
We build evaluation sets out of your real cases, including the awkward ones, and score the thing against them. Change a prompt or move to a new model and the suite runs again. You get a number for how often it is right, not a feeling.
Without that you cannot safely change anything, which is how AI projects quietly freeze.
Consultancy
Which processes pay back, and which do not. We will tell you when the answer is none of them.
Whether to build or buy. Plenty of problems are solved by software that already exists, and we are not short of work.
Model and cost engineering. The gap between a small model and a frontier one on a high volume workload is most of the running cost, and a lot of the job is deciding what needs the expensive one.
Governance. Approval checkpoints, audit trails, acceptable use for staff, and where you stand under the EU AI Act and ICO guidance if you are in scope.
Reviewing what you are being sold. If someone has quoted you for AI and you want a second opinion, that is a short piece of work and often the most useful one.
After it goes live
Building an agent takes a fortnight. Keeping it working is the rest of its life.
Models get retired, and every agent built on one has to be retested behaviour by behaviour and usually retuned. As a partner we see what is coming sooner, so it becomes planned work. Your rules change too, and that is a change to the agent, not a defect in it.
Common questions
Does being an OpenAI partner mean you only use OpenAI models?
No. It means we know the platform well enough to say when it is the right answer and when it is not. Where a client is already on AWS, or needs the model running inside their own account under their own controls, we use Bedrock, which carries OpenAI's open-weight models alongside Anthropic, Meta and Mistral behind one interface. Changing model later then does not mean rewriting the application.
Can you run models inside our own AWS account?
Yes. AWS is the infrastructure we ship on, and Bedrock lets the model run in your account under your own controls rather than ours. That matters most in regulated settings, where you hold somebody else's data under contract, or at volumes where the economics of model choice change. It is a different build from a straight API integration, so it is worth deciding before the architecture is set.
How do you know an AI agent is working properly?
We build evaluation sets out of your real cases, including the awkward ones, and score the system against them. Change a prompt or move to a different model and the suite runs again. That gives you a number for how often it is right, not an impression. Without it you cannot safely change anything, which is how AI projects quietly freeze.
What happens when the model an agent was built on is retired?
Every agent built on it has to be retested behaviour by behaviour against the replacement and usually retuned. As an OpenAI Select Partner we see what is coming sooner, which turns that from a scramble into planned work. It is covered by the monthly price rather than billed as a change, because it is maintenance, not something you asked for.